- torch
- cv2
- numpy
Train the model:
$ python3 src/train_model.py
Evaluate the model with mouse input:
$ python3 src/evaluate_model.py
There are 2 pretrained models saved in /src/models. The first model takes inspiration from this tutorial, here is the network:
The second model adds a new convolutional layer with 128 filters, a relu activation and a max pooling at the end of it.
When evaluation the model input mouse data, the image is take using opencv using a 500x500 canvas. Then the image is centered and finally is reduced to a 28x28 image (MNIST origianl image format).
Performances seem to be better when the input images are well centered.
Here is a link to the live process recorded on YouTube: